What Came Next
The problem this paper left on the table
Turing had described a destination. He had drawn the target on the wall. But he had not told anyone how to get there.
The Turing Test says: a machine can be called intelligent if it can talk like a human. But how does a machine learn to talk like a human? How does it learn anything at all?
In 1950, there was no answer. Computers could be programmed with explicit rules — if the user says X, say Y. But this approach, which we saw in our code section, breaks down immediately in the face of real conversation. There are too many possible inputs. You cannot write a rule for every sentence a human might type.
What was needed was a machine that could learn — that could look at examples of intelligence and extract the patterns itself, without being told every rule explicitly.
The unsolved problem Turing pointed toward
Late in his paper, Turing made a suggestion that turned out to be prophetic. He asked: instead of building an adult mind from scratch, why not build a child mind and educate it?
He wrote:
“Instead of trying to produce a programme to simulate the adult mind, why not rather try to produce one which simulates the child’s? If this is then subjected to an appropriate course of education, one would obtain the adult brain.”
This is, in essence, the idea of machine learning — the idea that a machine can learn from data and experience, rather than having every rule written in by hand.
Turing did not develop this idea into a working system. He died in 1954, too early to see what his suggestion would become.
But eight years after this paper, another researcher took the next step.
The researcher who built on this work
Frank Rosenblatt at the Cornell Aeronautical Laboratory read the same questions Turing had raised. He was not a philosopher — he was an experimental psychologist, inspired by how neurons in the human brain fire and learn.
In 1958, he built the Perceptron — the first machine that could actually learn from examples. It was trained on images and could learn to classify them correctly. Not by following rules someone had written. By adjusting its own internal weights based on whether it got the right answer.
This was the first working demonstration of machine learning. It was primitive — it could only learn simple patterns, and Marvin Minsky would later show it could not solve the XOR problem. But it was the first mechanical proof that a machine could learn, rather than just follow.
The Perceptron directly led to:
- Multi-layer neural networks
- Backpropagation (which made them trainable)
- And eventually, the transformer architecture behind ChatGPT and Claude
Every step in that chain begins with Rosenblatt’s attempt to answer the question Turing left open: not can a machine think, but how does it learn to?
Other directions from here
Not everyone followed Rosenblatt’s path. Some researchers tried a different approach: instead of learning from data, write down all of human knowledge explicitly in logical rules. This became known as symbolic AI or Good Old-Fashioned AI (GOFAI). It dominated from the 1950s through the 1980s, and then largely failed — the world turned out to be too messy and too large for explicit rules to capture.
Others focused on specific narrow problems: playing chess, proving mathematical theorems, recognising speech. These “narrow AI” systems scored important victories, but none of them looked like the general, flexible intelligence Turing had imagined.
It took until the 2010s, with the rise of deep learning, massive datasets, and the compute power to train enormous models, for something approaching Turing’s vision to become real.
That story spans the remaining 23 papers on this site.
Next paper: The Perceptron (1958) →
Frank Rosenblatt builds the first machine that learns from experience — the grandfather of every neural network in existence today.